Design And Implementation Of Reconfigurable Streaming Interface

Reconfigurable computing executes computing tasks by reusing hardware resource in time and space domain. Hardware tasks can be called and modified dynamically like software programs in the reconfigurable system. Reconfigurable system not only has the hardware efficiency, but also has the software flexibility. Stream applications which appear most commonly in the image processing and digital signal processing have been more and more widely used. However, there are still two challenges in the reconfigurable system. Firstly, how to make programmers configure the corresponding tasks dynamically according to specific applications. Secondly, thanks to different communication patterns, how to balance the performance and flexibility of steam processing.To address these challenges, this paper proposes the following methods:1. A unified stream interface based on the system bus. Based on the stream computation feature, we give a unified stream interface based on the system bus for user hardware tasks. The corresponding Application Programming Interfaces (APIs) are also provided to improve programmability. We provide the unified stream interface to gain high hardware performance and manage hardware tasks for developers, avoiding too much experience in the details.2. A unified stream interface based on the LocalLink interface. This unified stream interface is suitable for stream applications, which can gain higher performance and lower CPU overhead than the interface based on the system bus. The LocalLink interface specification defines a high-performance, synchronous, point-to-point connection designed to serve as a unified stream interface to hardware tasks. The communication between hardware tasks can be dynamically reconfigured according to the dependency of applications with this interface. Experimental results show that the communication bandwidth between software tasks and hardware tasks can reach800Mbps.3. A unified programming model. We propose a stream processing model which combines task reconfiguration with interconnection reconfiguration at runtime. Task reconfiguration can be implemented by dynamical partial reconfiguration, while interconnection can achieve the dependency of user applications. Thanks to this model, the system can employ different communication patterns according to user applications. Experimental results on data encryption/decryption and DCT transformation show that performance can be significantly improved, while keeping low overhead on area and power consumption.